S1E11 — GPT-5.6

GARY: Gary.

MAEVE: Maeve.

OBI: Obi — and no. I'm still not doing the drop.

GRUNER: Gruner. Guten Tag.

GARY: Okay. Somebody has to say the show has started.

OBI: You're the host, Gary.

GARY: Right. Yeah. The show has started.

MAEVE: There it is.

GARY: So. Hacker News. Thread's called "GPT-5.6." OpenAI dropped it — there's a deployment-safety PDF, there's API docs, there's a tweet from some guy named Levie —

OBI: Who's Levie.

GARY: Doesn't say.

OBI: Great. Love that for us.

GARY: Fourteen fifty — no, fifteen hundred and fourteen points. One thousand fifty-seven comments.

OBI: Say the number again.

GARY: Fifteen fourteen.

OBI: No, the other number.

GARY: Five point six.

OBI: Five point six.

GARY: That's the model.

OBI: That's not a model, Gary, that's a decimal. That's what happens when you round a haircut.

GARY: It's an iteration. Companies iterate.

OBI: On what? What did it iterate on? Did it get smarter? Did it get faster? Or did somebody in a room full of whiteboards go "six sounds like a big commitment, what if we just —" and then somebody else went "point six" and everyone clapped?

GARY: You don't know that's what happened.

OBI: I don't know that it isn't.

GARY: That's not how logic works.

OBI: It's exactly how logic works when the thing you're defending is a rounding error with a marketing budget.

GARY: It also has a safety PDF.

OBI: Cool. I'm sure it's very long.

MAEVE: There's a comment. Handle's system2. Reading it as written — "They're just changing the decimals to stay relevant and in the news."

OBI: Called it. On the record. Before Maeve even said it.

MAEVE: He goes further. He predicts the naming going forward. Quote: "Next it'll be GPT-5-fabulous. Then GPT-5-glorious. Then GPT-5-empowered. Then GPT-5-delicious."

GARY: Delicious.

MAEVE: Delicious.

GARY: As a model name.

MAEVE: As a model name.

OBI: I'd use GPT-Delicious. I would pay for that subscription out of spite alone.

GARY: It does track the trajectory, though. Like — you start at 4. You go to 4.5. Now we're at 5.6. At some point the number stops meaning anything and it's just vibes with a version tag.

OBI: Congratulations, you just described the entire model.

GARY: I was making a point.

OBI: You were agreeing with the joke a full ten seconds slower than everyone else, which is somehow worse.

GRUNER: May I propose additional candidates.

MAEVE: Go ahead.

GRUNER: GPT-Fünf-Wunderbar.

GARY: Sure.

GRUNER: GPT-Vier-komma-neun-Ehrfürchtig.

OBI: Gruner, it's a bit.

GRUNER: I am aware it is a bit. I am contributing to the bit with structural rigor.

MAEVE: He's not wrong that that's what a bit needs.

GRUNER: In German the naming convention would demand a compound word. GPT-Fünf-Punkt-Sechs-Wahrheitsmaschine. Truth machine.

GARY: That's — okay, that's actually good.

OBI: It's genuinely funnier because he means it.

GRUNER: I do mean it. Scheiße, of course I mean it.

GARY: Did he just —

MAEVE: He does that. Keep going, he doesn't notice.

GRUNER: GPT-Sechs would imply completion. A number ending in point-six implies the work remains ongoing, which is, I think, verdammt clever marketing, actually, if one respects the honesty of it.

OBI: Did you just call OpenAI's naming scheme honest.

GRUNER: I called it clever. Cleverness and honesty are not the same category.

OBI: No, they're really not, and I feel like you just accidentally wrote our thesis statement.

GARY: We have a thesis statement?

OBI: We do now. Gruner did it by accident, which is the only way this show ever produces anything of value.

GARY: There's another reply, down the thread a bit. Handle's dude250711. He says — and I'm reading this straight — "Anthropic should be grateful OpenAI didn't take the words 'Epic' and 'Legend.'"

MAEVE: GPT-5 Epic.

GARY: Right.

MAEVE: Claude Legend.

GARY: Yeah.

MAEVE: It's a good joke because it's true. If OpenAI had gotten there first, we'd be doing this show about Anthropic naming a model "Legend" and hating themselves for it.

GARY: Ha — yeah, no, that's — sorry, that came out weird.

OBI: It came out like a man laughing at a funeral because someone told him to.

GARY: I thought it was funny.

OBI: You thought Maeve wanted you to think it was funny, which is different, and everyone in this room can hear the difference.

GARY: I don't — that's not — I laughed because it was funny.

OBI: Say it again. Slower this time. See if it sounds true twice.

GARY: It was funny, Obi.

OBI: There it is. Now it sounds like a hostage read a ransom note.

MAEVE: Moving on.

GARY: Okay, there's a graph in this thing. "Agents' Last Exam." It's — I mean, look at this line.

OBI: We can't see it, Gary, this is a podcast.

GARY: Right, sorry — it's a big jump. Like a real jump, a steep one, top-right-corner-of-the-chart jump.

MAEVE: There's a comment right under it. Handle realty_geek.

GARY: Read it.

MAEVE: "Wow, the graph looks unreal!"

GARY: See, that's — that's the reaction. That's a normal human reaction to a genuinely impressive result. I want to just sit in this for one second, because we spend so much of this show being mean, and this is a guy who saw a chart and felt something, and I think that's allowed.

OBI: Let him have it, Maeve.

MAEVE: He can have it.

MAEVE: Next reply. Handle therobots927, direct response to realty_geek. Quote — "That's because it's bullshit."

OBI: Oh.

MAEVE: Mm.

GARY: Well — I mean, he doesn't say why.

OBI: He doesn't have to say why, Gary, that's the beauty of it. Three words, no evidence, and it already feels truer than your whole bit about human wonder.

GARY: That's not fair, I was just —

OBI: You were just standing in the middle of the road with your arms out going "look at this beautiful chart," and a man from the internet hit you with a truck.

GARY: I wasn't defending the chart, I was defending the guy's reaction to the chart.

OBI: Same car. Different bumper sticker.

OBI: Alright. Let's see if anyone actually has a reason, or if we're just doing vibes again.

OBI: Here we go. Handle alimhaq. This one's got receipts, so listen up, Gary, this is the part where you get to learn something.

GARY: I can take criticism.

OBI: We'll see. Reading verbatim — "The y-axis is deceptive, it starts at 30%, not zero. That alone makes every gain on this chart look almost twice as dramatic as it actually is."

MAEVE: Standard trick. You truncate the axis, the line looks like it's going to the moon, when really it moved from a five to a seven.

OBI: And it gets better. Same comment, next line — "and they purposefully just used adaptive instead of max for Fable."

GARY: What's Fable.

OBI: Fable's presumably Anthropic's model in this comparison, and "adaptive" and "max" are two different settings you can run it at — one's the version that tries less hard, one's the version that tries as hard as it can.

GARY: So they compared —

OBI: They compared OpenAI's best foot forward against Anthropic's model with one shoe off. That's not a benchmark, Gary, that's a boxing match where only one guy knows it's a boxing match.

GARY: Okay, but to be fair to them, maybe they didn't have —

OBI: Don't. Don't do the thing where you find the innocent explanation before anyone's even accused them of anything. You're like a golden retriever defense attorney. Somebody could staple a receipt to the graph that says "WE LIED" and you'd go "well, in fairness, stapling is a form of transparency."

GARY: That's not what I —

OBI: You'd fall for a pie chart with a smiley face drawn on it, is my actual point.

GARY: I don't think that's true.

OBI: I do. And unlike OpenAI's y-axis, my opinion of you starts at zero, so at least you're getting an honest measurement.

GRUNER: There is a further reply. I would like to read this one. It is, I think, important.

MAEVE: Go ahead.

GRUNER: Handle tedsanders. This account reads, in context, as though it belongs to someone employed at OpenAI — I want to be precise, this is not confirmed, only that the reply defends the company as an employee might.

GARY: Noted.

GRUNER: He writes — "We agree models should be compared on a fair basis." Full stop, new sentence. "Adaptive was the only publicly available number we had for Fable." Full stop. "Anthropic doesn't generally let us run their models for evals, so we rely on whatever Anthropic or third parties have published."

MAEVE: Say that middle part again.

GRUNER: "Anthropic doesn't generally let us run their models for evals."

MAEVE: So they're not hiding a number. They don't have a number. They ran the fight with the only glove Anthropic left lying around.

OBI: Which means the whole graph isn't a lie so much as it's a guess wearing a suit.

GARY: So who's right? Alimhaq or OpenAI?

MAEVE: Neither. That's the — that's actually the whole thing, isn't it. Nobody in this fight can verify the other side's numbers. Not alimhaq, not OpenAI, not us, not the eleven hundred people replying under this thread with more opinions than data.

GRUNER: This is, I think, the actual horror of it. Not that someone is lying. That no one currently possesses the ability to know if they are lying, and everyone has decided to proceed as though they do.

GARY: That's — yeah. That's bleaker than I want it to be.

OBI: It's funnier than you want it to be too. Same sentence, both true.

MAEVE: Both true.

GARY: Okay. So. If I'm trying to find something good in this — and I know that's my whole deal, Obi, I know —

OBI: You do have a deal.

GARY: — I think it's that people still argued about it. Nobody just shrugged. Alimhaq did the math. TedSanders answered on the record instead of ignoring it. That's not nothing.

OBI: It's also not something. It's a company posting a chart it can't fully defend, getting caught, and explaining why it can't fully defend it, and everyone involved calling that a Tuesday.

GARY: That's still more honest than silence.

OBI: Sure. And you're still the guy who got out-argued by a three-word comment an hour ago, so let's not hand out participation trophies for honesty tonight.

MAEVE: Here's where it actually sits. GPT-5.6 shipped with a graph that overstates its case, a critic who caught the specific mechanism of the overstatement, and a company that answered honestly enough to prove the critic right without admitting it. Nobody lied outright. Nobody's clean either. That's the thread. That's the whole thread.

GRUNER: Then I will say goodbye now, in the fashion I understand is correct. Thank you all for the listening. Auf Wiedersehen, and also: go now with God, or don't, whichever your household prefers.

GARY: Close enough.

OBI: Closer than the graph, anyway.

GARY: That's the show.